Generation capacity expansion under long-term uncertainties in

the US electric market

Alexandre Klein1 , Julian Bouchard2 , Sabine Goutier2

1 EDF R&D and Paris Dauphine University 1 , Phone +33147653687, E-mail:

alexandre.klein@edf.fr
2 EDF R&D2

Abstract
In this paper, we deal with generation capacity expansion under long-term uncertainties
regarding fuel prices and CO2 emissions regulation. We present a model based on stochastic
dynamic programming which gives optimal generation investment planning for perfectly
competitive power markets. It is applied to the US continuous electricity market with
DOE's fuel price scenarios. We show that taking into account uncertainties on fuel costs
and on CO2 emissions regulation can change the optimal investment decisions. Moreover,
we show that, for generators, the level of risk-aversion is a major factor inuencing the
investment decisions.

1 Introduction

The recent high volatility in fuel markets, combined with environmental regulation policies, has
introduced major uncertainties into the planning of generation capacity expansion. These uncer-
tainties make generators' decisions to invest in new capacities more dicult. The literature has
focused mainly on long-term demand uncertainties, but little has been done regarding fuel price
and environmental policy uncertainties. This article discusses the optimal generation investment
choices made in an electricity market with fuel price and regulatory uncertainties over the period
2010-2030. It focuses particularly on long-term uncertainties surrounding coal and gas prices,
and on CO2 emissions reduction policy. We have developed an optimization model for electric
generation investments based on stochastic dynamic programming to tackle this issue.
To illustrate our discussion, we consider the case of the US long-term electricity market seen as
perfectly competitive. The electricity market data and the fuel price scenarios are taken from
the Annual Energy Outlook 2008 of the US Department of Energy (DOE).

The paper is organized as follows: in the next section, is summarized the related literature.
Then, in section 3 is presented the generation investment model. Section 4 describes the general
data and assumptions made regarding the long-term US electricity market and section 5 presents
the results obtained for this market. Finally, section 6 concludes.

2 Related literature

As far as electric generation capacity expansion is concerned, many articles in recent years
have dealt with the deregulation of the electricity market in comparison with the former public
monopoly (Murphy and Smeers [1] for example). But application of real option theory to
1†
EDF R&D, 1 Avenue du Général de Gaulle, F-92141 Clamart Cedex
& Finance of Energy Markets Laboratory, joint laboratory Paris Dauphine University, CREST and EDF R&D
2†
EDF R&D, 1 Avenue du Général de Gaulle, F-92141 Clamart Cedex

1
electricity markets gave rise to a large literature on the impact of long-term uncertainties on
generation investment decisions. Most of these papers consider a perfectly competitive market
where investors are price-takers and face an uncertain future. For example, Deng et al. [2] value
an investment in thermal power plant in a market where electricity and gas prices are seen as
stochastic future contracts (geometric Brownian motion or mean-reverting process). However,
the price-taker assumption is questionable in electricity markets because generation investments
will modify electricity prices in the long-term since costs and technological mix are changed.
In order to take into account the eect of new generation investments on electricity market
price, some papers consider models where electricity prices are endogenous. Pineau and Murto
[3] developed a dynamic stochastic oligopoly model applied to the Finnish electricity market.
An uncertain demand is introduced and is modeled as a markovian event tree. Although based
on an open loop information structure, the model helps to understand dynamics of production,
investment and market power in a medium time horizon. One of their results shows that the
uncertain demand can lead players to delay their investments in order to gain some information
on the future demand. Botterud et al. [4] use a similar representation of an uncertain demand in
their capacity expansion model. They focus mainly on the impact of the market design (energy-
only market, capacity payments) on the generation investment decisions both on a centralized
and a decentralized electricity market (represented with a single prot-maximization player).
Then a large part of this literature considers only uncertainty on the demand (the main challenge
is to guarantee load supply). But little has been done regarding the impact of uncertainties on
fuel prices and on CO2 regulation on generation capacity expansion.

3 The generation investment model

Assuming that the long-term electricity market is perfectly competitive, the model determines
the optimal investment decisions in order to maximize the social welfare over a time period. A
backwards stochastic dynamic programming (SDP) algorithm is used to solve the investment
problem. The model uses an annual time step. At each time step, a set of new investments
can be decided which takes place at the beginning of each time step. No construction delay is
modeled here. Two types of technologies are considered:
• the initial technologies: they can be decommissioned over the period. No new investment
in these technologies will be made
• the new technologies: generators can invest in a set of new technologies. The quantity a
generator can invest in is not a continuous variable but a discrete capacity step. Once
some generators have decided to invest in a new technology, it will not be decommissioned
over the period

From Soyster and Murphy [5], we know that a capacity expansion model designed for an
ecient public monopoly converts directly into one applicable to a perfectly competitive market.
Then, the overall problem is stated as the following optimization problem:
T
X −1
gt (lt , it , xt )(1 + r)−t + gT (lT , iT , xT )(1 + r)−T
 
min ES
(it )t∈T
t=1
s.t xt+1 = xt + it − dt

2
where:
T planning horizon (years)
t time step (year)
S the set of long-term scenarios
it vector of invested capacities (MW) - command variable
dt vector of decommissioned capacities (MW)
xt vector of existing capacities (MW) - state variable
lt load duration curve (MW, hourly block)
r risk-adjusted real discount rate
gt objective function-total cost function ($)
gT terminal cost function($), year T

The total cost function at time t is gt (lt , it , xt ) = CtGen (lt , xt ) + CtO&M (xt ) + CtInv (it ) where:
• CtGen (lt , xt ) is the generation cost. For each technology, we can convert the fuel price as-
sociated and the CO2 price into a generation variable cost ($/MWhr). Then, the model
computes a merit order stack on the generation variable cost in order to minimize genera-
tion costs. If the demand load is curtailed, then the electricity market must pay the value
of lost load (VOLL) proportionally to the curtailed energy
• CtO&M (xt ) is the Operation and Maintenance costs. For each technology O&M costs are
dened. They correspond to xed yearly costs ($/MW.year). The model calculates the
total O&M costs for the electric system
• CtInv (it ) is the investment cost. We associate to each new technology a xed investment
cost ($/kW) and a lifetime (years). Then, the model calculates constant annuities for new
investments

Generally, SDP assumes the terminal values being the nal values of the objective function
associated to the state variables. Addressing terminal values can be intricate since they corre-
spond to the end of the period whereas there is no information on the future. In our modeling,
we have not such a problem since the investment cost is annualized. Then, the terminal cost
function gT (lT , iT , xT ) can therefore simply be set equal to the cost function in the last period.

We consider a possible uncertain future on fuel prices, CO2 taxation and demand growth.
These uncertainties are modeled as a discrete markovian event tree denoted S . At each time
step, the system can be in dierent states and transitions from one state to another occur with
given probabilities (see Figure 1).

Figure 1: Event tree detail

Finally, the backward problem is solved using the Bellman algorithm. It consists in calculat-
ing backwardly the investment decisions that minimize the objective function (here the total cost

3
function). For each node n of the event tree, for each possible vector of the state variable (here
the existing capacities) xt , the algorithm calculates a Bellman value V bn (xt ) which corresponds
to the minimum of the expected future objective function at node n:
 
X
V bn (xt ) = minn(it ) gt (lt , it , xt )(1 + r)−t + proban0 /n V bn0 (xt + it )
t∈T
n0 ∈F (n)

where n0 is a node son of n, F (n) is the set of all the nodes son of n and proban0 /n is the
probability to switch from node n to node n0 .

4 Case study from the US electricity market

4.1 The general problem

We analyse potential investments in new generation capacities on the US electricity market

over the period 2010-2030. Investments are decided every ve years and plants start producing
immediately. We make strong assumptions on decommissioned capacities since we are more
interested in determining the long-term optimal generation eet rather than forecasting the
future real investments. The US electricity market is seen as a perfectly competitive market.
We use the three fuel price scenarios from the Annual Energy Outlook 2008 [6] : the Reference,
the High price and the Low price cases. First, we describe the optimal investment decisions if
the Reference case occurs. The impact of dierent CO2 prices on the generation mix will be
studied. Then, we combine the three scenarios in an event tree centered around the Reference
Price scenario with constant probabilities to switch denitely to the Low or to the High price
scenarios. Assuming generators are risk-neutral, we test the robustness of the Reference case
investment decisions to these uncertainties. Finally, we assume that generators are no longer
risk-neutral but are risk-averse. We compare the investment decisions with the previous risk-
neutral case.

4.2 General data on the US electricity market

The following data are taken from the Annual Energy Outlook 2008 [6] and from the FERC [7].
The generators can invest either in new coal plants, in new combined-cycle gas power plants
(CCGTs) or in new oil peak plants. Investments in new nuclear plants will be analysed in a
specic section. Table 1 presents the technical specications of the new power plants.
Coal plant CCGT Nuclear plant Peak unit
Heat rate BTU/kWhr 9200 7196 10400 10833
Investment cost$/kW 1534 717 2475 500
Fixed O&M $/kW.year 26.8 12.1 66.1 11.8
Variable O&M $/MWhr 4.5 2 0.5 3.5
CO2 emissions ton/MWhr 0.74 0.35 0 0.83
Lifetime years 35 25 40 25
Size MW 600 250 1350 160

Table 1: Technical specications

The initial generation eet capacity in 2005 equals 905 GW. Details on the initial capacities
are given in Table 2.

4
 Coal plants 305 GW
Gas plants/CCGTs 258 GW
Peak units 127 GW
Nuclear plants 100 GW
Renewable energies 15 GW
Pumped storage 22 GW
Conventional Hydro power 77 GW
Total 905 GW

Table 2: 2005 US generation eet

We make the following assumption regarding the power plant decommissionings: 20% of all
initial technologies are phased-out every ve years (except for renewable energies and hydro-units
which are not phased-out or replaced identically). In order to consider only economic trade-os
between the technologies, we set the load factor to 90% for all the thermal technologies and to
25% for the renewable energies.

The fuel price scenarios for coal (Average delivered price), gas (Henry Hub price) and oil
(Distillate fuel oil) are presented in Figure 4 (at the end of the paper). These scenarios are
taken from the Annual Energy Outlook 2008 [6]. Then, using the technical specications of the
technologies, these fuel prices are converted into variable generation costs for each technology.
In addition to fuel variables, we add a CO2 emission price. Like on the EU CO2 market, it
represents the emission cost for producing 1 MWhr of electricity.
We combine these three scenarios as an event tree centered around the Reference scenario with
constant probabilities PLow and PHigh to switch denitively to the Low or High price scenarios.
The switch can occur every ve years. This event tree is presented in Figure 2.

Figure 2: Event tree from the AEO 2008 scenarios

We assume that the US electric grid is perfectly interconnected. The US continuous load
duration curve was made using FERC data [7] by aggregating the regional electric grids load
duration curves in 2005 (Figure 3). We assume that the electricity demand growth will be sup-
plied by investments in renewable energies, so we do not consider any demand growth over the
period.

5
 Figure 3: 2005 Continuous US Load duration curve

Finally, we choose a discount rate equal to 8% and we set the VOLL to 20 000 $/MWh.

5 Results

5.1 Result 1: Optimal investment planning for the Reference scenario - no

nuclear technology

In the Reference scenario, with no CO2 price, we observe that the optimal investment decision
is a mix between coal units, CCGTs and peak units. Finally, a total of 750 GW are installed
over the period, mainly in coal units (510 MW). The coal plants are the most competitive to
supply base load while CCGTs are more ecient to supply semi-base load. Finally, 80 GW are
installed to supply peak-load.
GW 2010 2015 2020 2025 2030
Coal units 0 30 160 160 160
CCGTs 40 120 0 0 0
Peak units 80 0 0 0 0

Table 3: Reference scenario investment decisions - no nuclear and CO2 price=0 $/ton

Then, we test the sensitivity of investment decisions to dierent CO2 prices (each price is
supposed constant over the period). For a CO2 price inferior to 40 $/ton, the technology mix
is still made of coal units to supply base-load and of CCGTs to supply semi-base load. Above
40 $/ton, CCGT becomes the most competitive technology both for base and semi-base load.

In the following (results 2 and 3), we will use a CO2 price equal to 30 $/ton. Therefore, we
present in table 5 the optimal investment planning for further comparisons.

6
 CO2 price ($/ton)
GW
10 20 30 40 50
Coal units 480 450 400 160 0
CCGTs 190 220 270 510 670

Table 4: Sensitivity to CO2 price

GW 2010 2015 2020 2025 2030
Coal units 0 0 80 160 160
CCGTs 40 150 80 0 0
Peak units 80 0 0 0 0

Table 5: Reference scenario investment decisions - no nuclear and CO2 price=30 $/ton

5.2 Result 2a: Impact of fuel price uncertainties on the investment decisions
- no nuclear technology

We set CO2 price to 30 $/ton over the period. Uncertainties are modeled as the event tree
presented in Figure 2. We consider successively the possibility to switch to the Low and to the
High scenario with dierent probabilities PLow and PHigh (if PLow > 0 then PHigh = 0 and
inversely). We focus on the investment decisions if the Reference scenario occurs.
Reference
GW Low
40% 30%
PLow
20% 10% (CO2 30$/ton) 10%
PHigh
20% 30% 40%
High
Coal units 0 160 320 370 380 400 410 440 450 450 480
CCGTs 670 510 350 300 290 270 260 230 220 220 190

Table 6: Impact of fuel price uncertainties - no nuclear technology

The results show that the optimal investment mixes in the Reference and High price scenar-
ios are fairly similar (coal units predominant). However, if the Reference scenario occurs, we
observe that the higher PHigh is, the higher the investments in coal units are. A high probabil-
ity to switch to the High scenario leads generators to increase their investments in coal units in
order to hedge against the risk of high gas prices. But, since the overall fractions in the dierent
technologies are not changed, we can conclude that the investment decisions obtained in the
Reference scenario are robust to uncertainty to switch to the High price scenario.
Conversely, we cannot conclude on the robustness of the investment decisions for the Refer-
ence scenario as far as uncertainties to switch to the Low scenario are concerned. Indeed, the
decisions if the Low scenario occurs lead to invest only in CCGTs. Then, for high probability
to switch to the Low scenario, we show that CCGTs become predominant (especially if PLow is
greater than 30%).
Moreover, investments in peak units are constant for all the cases and are equal to 80 GW
over the period.

7
5.3 Result 2b: Impact of fuel price uncertainties on the investment decisions
- with nuclear technology

We consider the possibility for the generators to invest in nuclear plants in addition to CCGTs,
coal and peak units. We test if the investment decisions for the Reference scenario are identically
modied by fuel price uncertainties as previously in Results 2a.

Reference
GW Low
40% 30%
PLow
20% 10% (CO2 30$/ton) 10% 20%
PHigh
30% 40%
High
Coal units 0 0 0 0 0 0 0 0 0 0 0
CCGTs 200 200 170 170 170 170 170 170 170 170 170
Nuclear units 470 470 500 500 500 500 500 500 500 500 500

Table 7: Impact of fuel price uncertainties - with nuclear technology

We show that, whatever scenario we consider, generators no longer invest in coal units but
in nuclear units in order to supply base-load demand. CCGTs are the most competitive to serve
semi-base load. As previously, peak unit investments are constant and equal to 80 GW.
The most striking point is that, if nuclear investments are allowed, the investment decisions
in the Reference scenario are robust to the risk to switch to another scenario. Therefore, we
can conclude that considering possible nuclear investments make the decisions not impacted by
uncertainties regarding fuel prices.

5.4 Result 3: Impact of risk-aversion on the investment decisions - no nuclear

technology

Previously, generators took their investment decisions considering that each scenario had a
probability to occur and solved the problem using mathematical expectation. In this section,
we assume that the generators are risk-averse: they take their investment decisions in order to
minimize the maximum cost over the period. This criterion is known as the 'MinMax Regret'
criterion. Then, the optimization program the generators have to solve turns into :
T −1
!
X
−t −T
min max gt (lt , it , xt )(1 + r) + gT (lT , iT , xT )(1 + r)
(it )t∈T S
t=1

s.t xt+1 = xt + it − dt
The objective of this section is to see what is the worst fuel and CO2 price trajectory
that risk-averse generators will focus on in their decisions in the uncertain future modeled as
previously (Figure 2). Then, if we focus on the Reference scenario, generators will take their
investment decisions considering the possibility to switch either to the high price or to the low
price scenario.
We show that the investments are similar to the ones made in the deterministic High price
scenario. The worst case for generators corresponds to the scenario with relative high gas prices.
Then, their best strategy in order to minimize the maximum total generation cost is to focus
only on the High price scenario.

8
 GW 2010 2015 2020 2025 2030
Coal units 0 0 160 160 160
CCGTs 40 150 0 0 0
Peak units 80 0 0 0 0

Table 8: investment decisions for risk-averse generators - CO2 price = 30 $/ton

6 Conclusions

We show that generation capacity expansion planning in the continuous US is sensitive to un-
certainties regarding fuel and CO2 costs as far as nuclear unit investments are not considered.
Indeed, if only fossil thermal units are allowed, the DOE Reference scenario is not robust to
these uncertainties. Conversely, allowing nuclear investments make the generation investment
decisions robust. Finally, we show that generators react dierently to uncertainties on fuel and
CO2 costs depending on risk-aversion.

References

[1] FH.Murphy and Y.Smeers, Generation Capacity Expansion in Imperfectly Competitive Re-
structured Electricity Markets, Operations Research, 53, 646-661, 2005

[2] SJ.Deng, B.Johnson and A.Sogomonian, Exotic electricity options and the valuation of elec-
tricity generation and transmission assets, Decis Support Syst, 30, 383-392, 2001

[3] PO.Pineau and P.Murto, An Oligopolistic Investment Model of the Finnish Electricity Mar-
ket, Annals of Operations Research, 121, 123-148, 2003

[4] A.Botterud, M.D.Ilic and I.Wangensteen, Optimal investments in power generation under
centralized and decentralized decision making, Power Systems IEEE Transactions on, 20,
254-163, 2005
[5] FH.Murphy and AL.Soyster, Economic behavior of electric utilities, Prentice-Hall Englewood
Clis NJ, 1983
[6] US Department of Energy /EIA, Annual Energy Outlook, 2008
[7] Federal Energy Regulatory Commission, Form FERC-714 data, 2005
[8] US Department of Energy /EIA, Uranium Marketing Annual Report, 2007

9
Figure 4: Annual Energy Outlook 2008's fuel price scenarios

10